I agree that stronger, more nuanced interpretability techniques should tell you more. But, when you see something like, e.g.,
25132 ▁vs, ▁differently, ▁compared, ▁greater, all, ▁per
25134 ▁I, ▁My, I, ▁personally
isn't it pretty obvious what those two autoencoder neurons were each doing?
Epistemic status: Half-baked thought.
Say you wanted to formalize the concepts of "inside and outside views" to some degree. You might say that your inside view is a Bayes net or joint conditional probability distribution—this mathematical object formalizes your prior.
Unlike your inside view, your outside view consists of forms of deferring to outside experts. The Bayes nets that inform their thinking are sealed away, and you can't inspect these. You can ask outside experts to explain their arguments, but there's an interaction cost associated with inspecting the experts' views. Realistically, you never fully internalize an outside expert's Bayes net.
Crucially, this means you can't update their Bayes net after conditioning on a new observation! Model outside experts as observed assertions (claiming whatever). These assertions are potentially correlated with other observations you make. But because you have little of the prior that informs those assertions, you can't update the prior when it's right (or wrong).
To the extent that it's expensive to theorize about outside experts' reasoning, the above model explains why you want to use and strengthen your inside view (instead of just deferring to outside really smart people). It's because your inside view will grow stronger with use, but your outside view won't.
I think that many (not all) of your above examples boil down to optimizing for legibility rather than optimizing for goodness. People who hobnob instead of working quietly will get along with their bosses better than their quieter counterparts, yes. But a company of brown nosers will be less productive than a competitor company of quiet hardworking employees! So there's a cooperate/defect-dilemma here.
What that suggests, I think, is that you generally shouldn't immediately defect as hard as possible, with regard to optimizing for appearances. Play the prevailing local balance between optimizing-for-appearances and optimizing-for-outcomes that everyone around does, and try to not incrementally lower the level of org-wide cooperation. Try to eke that level of cooperation up, and set up incentives accordingly.
The ML models that now speak English, and are rapidly growing in world-transformative capability, happen to be called transformers.
Two moments of growing in mathematical maturity I remember vividly:
I found it distracting that all your examples were topical, anti-red-tribe coded events. That reminded me of
In Artificial Intelligence, and particularly in the domain of nonmonotonic reasoning, there’s a standard problem: “All Quakers are pacifists. All Republicans are not pacifists. Nixon is a Quaker and a Republican. Is Nixon a pacifist?”
What on Earth was the point of choosing this as an example? To rouse the political emotions of the readers and distract them from the main question? To make Republicans feel unwelcome in courses on Artificial Intelligence and discourage them from entering the field? (And no, I am not a Republican. Or a Democrat.)
Why would anyone pick such a distracting example to illustrate nonmonotonic reasoning? Probably because the author just couldn’t resist getting in a good, solid dig at those hated Greens. It feels so good to get in a hearty punch, y’know, it’s like trying to resist a chocolate cookie.
As with chocolate cookies, not everything that feels pleasurable is good for you.
That is, I felt reading this like there were tribal-status markers mixed in with your claims that didn't have to be there, and that struck me as defecting on a stay-non-politicized discourse norm.
2. The anchor of a major news network donates lots of money to organizations fighting against gay marriage, and in his spare time he writes editorials arguing that homosexuals are weakening the moral fabric of the country. The news network decides they disagree with this kind of behavior and fire the anchor.
a) This is acceptable; the news network is acting within their rights and according to their principles
b) This is outrageous; people should be judged on the quality of their work and not their political beliefs…
12. The principal of a private school is a member of Planned Parenthood and, off-duty, speaks out about contraception and the morning after pill. The board of the private school decides this is inappropriate given the school’s commitment to abstinence and moral education and asks the principal to stop these speaking engagements or step down from his position.
a) The school board is acting within its rights; they can insist on a principal who shares their values
b) The school board should back off; it’s none of their business what he does in his free time…
[Difference] of 0 to 3: You are an Object-Level Thinker. You decide difficult cases by trying to find the solution that makes the side you like win and the side you dislike lose in that particular situation.
[Difference] of 4 to 6: You are a Meta-Level Thinker. You decide difficult cases by trying to find general principles that can be applied evenhandedly regardless of which side you like or dislike.
--Scott Alexander, "The Slate Star Codex Political Spectrum Quiz"
Say there are two tribes. The tribes hold fundamentally different values, but they also model the world in different terms. Each thinks members of the other tribe are mistaken, and that some of their apparent value disagreement would be resolved if the others' mistakes were corrected.
Keeping this in mind, let's think about inter-tribe cooperation and defection.
In the worst equilibrium, actors from each tribe evaluate political questions in favor of their own tribe, against the outgroup. In their world model, this is to a great extent for the benefit of the outgroup members as well.
But this is a shitty regime to live under when it's done back to you too, so rival tribes can sometimes come together to implement an impartial judiciary. The natural way to do this is to have a judiciary classifier rule for reference classes of situations, and to have a separate impartial classifier sort situations into reference classes.
You're locally worse off this way, but are globally much better off.
Use your actual morals, not your model of your morals.